concise (abbreviated) taiwanese tone sandhi

i.e. there are, in fact, more possibilities than given in the initial list (as represented in the marginally disparate literature) but this list MIGHT be a good place to start to look for what to listen for.*

"tone sandhi" refers to tone shifts depending on whether an item is "phrase-final" or not.

My current teacher, Liu Yiru 劉怡如, following her teachers at Cheng Gong 成功, disputes this simple characterization, characterizing the situation as a more complex question of possible interactions between tones, and also suggests "more difficult" tones are more likely to change (just as mandarin 3-3s go 2-3). The diagrams on the sample pages would seem to support this contention, or see below for a list of some questionable examples.

NOTE: the initial list below reverses the typical presentation of sandhi which treats final, also "isolate", items as primary. I have done this in the belief that non-final contexts are more frequent. (that is certainly not equally true for all items) I presume frequency is a good guiding principle for pedagogy. (I also suspect phrase endings might be considered "more marked", and hypothesize that sandhi exists to "emphasize" phrase boundaries (though I'd better think alot harder about what kind of evidence can validate that sort of hypothesis...)) FURTHER NOTE: If it's more about tone interrelationships than simple final/non-final positioning, then this shifted focus may be misdirected, and prioritizing "isolates" before seeing how they interact in different contexts may be more convenient.

The following list summarizes the intersection of claims found in Samuel Wang 1992 and Mike Campbell's "Learning Taiwanese" page (at

(tone contours are flat unless otherwise indicated, "low" may be more accurately described as "mid")

non-final final
high high,fall
high,fall low,fall
low,fall low
low high
high,stop low,stop
low,stop high,stop
(just Wang:)
rise high,fall
(Wang also implies that EITHER low tone (flat or falling) could have a rising tone in final position) (Conversely, Campbell's system has NO non-final rising tones)

Both agree the stop tones simply reverse. The first four seem to show a fairly nice circular pattern, of high leading to falling leading to lower falling to low which then becomes high again.

I must STRESS again though: THESE ARE NOT ALL THE POSSIBLE SHIFTS which the original accounts suggest, i.e. that presumably one might hear.

Below, the original claims are summarized.

from Samuel Wang 1992

indicating pitch and contour with a double index where 5 is high and 1 is low
"isolation" "sandhi"
55 33
33 21
21 51
51 35(55)
24 21(33)
4 3(2)
3 4

from ("Mike Campbell")

tone numbering and descriptions:

1. high, flat
2. high, fall
3. low, fall
4. low, short (always cut-off by stop consonant)
5. low, rise
6. high, fall (=2 above)
7. low, flat
8. high, short (always cut-off by stop consonant)

(actually Campbell describes tone 7 as "low and flat" but his sandhi descriptions refer to it as "mid tone"; the latter correlates better with Wang's "33" notation, but i stick with "low" to perpetrate a binary reduction (for mnemonic purposes) while recognizing that empirical manifestations are so much more complex)
campbell's rule list wang's rules in campbell's notation campbell's rules in wang's notation
1 7
5 7
2 1
3 2
7 3
4 8
8 4
4 2
8 3
wang# campbell#
55 1
33 7
21 3
51 2,6
24(35) 5
4 8
3 4
1 7
7 3
3 2
2 5(1)
5 3(7)
8 4
4 8
55 33
24 33
51 55
21 51
33 21
3 4
4 3
3 51
4 21
The discrepancies are mostly in the multiple options. Campbell's 2(=w51) and 5(=w24) only include what Wang lists in parentheses. Meanwhile Campbell includes extra options for 4(=w3) and 8(=w4), the stop tones. Notice that Campbell's system suggests rising tones can only be heard phrase finally. (no c5 appears in the second, "sandhi" colomn) (for him, phrase-internal low tones manifest phrase-finally as either high or falling tones). Wang, on the other hand, prioritizes 51-35. 2 other accounts I've looked at agree with Campbell that rising tones do not occur phrase internally.

Wang also also prioritizes 24-21, while including 24-33. Simpson and Wu suggest that rising tones go low falling in northern Taiwan, but low flat in southern Taiwan.

Simpson and Wu clarify the c4,c8 options as flipping when ending in p/t/k, but undergoing their respective c2,c3 when ending in a glottal stop.

Other tone comments

Wi-vun Tiaffalo Chung cites Ong(1993) as suggesting the rising tone can actually be a falling-rising tone (as Simpson and Wu describe it), and that others suggest that tone 4 and 8 (the stop tones) are merging together.

In fact, that study ( provides evidence that these changes are occuring more in young people and in the North, and thus may be due to the influence of Mandarin's spread on the island.

(Though since Mandarin also has a rising tone, I find it a bit regrettable that Mand3 didn't creep into some other tone's place.)

(Also, in what sense would Mandarin, which lacks stop tones, cause a high and low tone to merge? Is it a step to their abolition? Notice they were already pretty close anyways. Is it just basic language simplification? for efficient learning, broader spreading, or lack of cultural support? I'd also like to get more clear on word by word changes that are occuring: are all 4 and 8 words taking on the same tone? or are just the tones losing the distinction, while some words could be resorting themselves into new tones, such as high and low, to maintain distinctiveness?)


*In fact, my next step is to look over the praat diagrams on the other pages to see how well these descriptions correlate with sandhi situations there. More recording may be necessary. In any case, pitch contours there already appear yet more complex. Still one must also keep in mind that praat's pitch inferencing is not infallible either.


random observations from the pictures

"eaten?" vs. "eat" almost gives a good example of high to highfalling, except "ba(飽)" in the former case in fact also seems to fall towards the mid tone, just not as much as when it is phrase-final

on the other hand, "haven't eaten" could suggest 3 distinct level tones (high, mid, and low), unless "gwa(我)" is some kind of highrising and/or "ja(吃)" was the lowfalling, the latter of which seems unlikely (though all its examples do have lil curls at the front). Of course, highrising isn't in traditional descriptions either. THAI, on the other hand, definitely has 3 level tones, though I am under the (hopefully correct) impression that the high tone can come with a slight rise, and the low tone with a slight fall. In any case, to follow traditional linguistic argumentation one would have to look for "minimal pairs", in fact, in this case, i guess, "minimal triads" (distinct items identical for all but the feature in question). [that could possibly raise the question if whether "linguistic logic" doesn't miss aspects of "fluent accent"]

"eat together?" vs. "eat rice" however seems to show the same low high relationship between 吃 and 飯, even though 飯 is not final in the former case. admittedly, it's nowhere near as high as the either high tones, so what is it supposed to be? is it more evidence for a mid tone? is there really no sandhi in these cases?

"oysters" vs. "oyster omelette" seems to give another good example of a high tone falling when phrase-final.